https://orcid.org/0000-0003-3523-3333
Figures
Abstract
Objectives
Linking neighborhood- and patient-level data provides valuable information about the influence of upstream social determinants of health (SDOH). However, sharing of these data across health systems presents challenges. We set out to develop a pipeline to acquire, deidentify, and share neighborhood-level SDOH data across multiple health systems.
Methods
We created a pipeline centered around Decentralized Geomarker Assessment for Multi-Site Studies (DeGAUSS) that utilizes containerization to geocode patient addresses and obtain neighborhood-level SDOH variables. We compared DeGAUSS to a third-party vendor geocoding tool available at Duke Health using a cohort of adult patients referred for abdominal transplant from January 1, 2016, to December 31, 2022. We calculated Cohen’s Kappa and percent disagreement at census block group and tract levels, and by Area Deprivation Index, urbanicity, and year.
Results
The pipeline successfully generated SDOH data for 97.8% of addresses. There was high concordance between DeGAUSS and the vendor tool at the census block group (0.93) and tract levels (0.95). At the block group level, disagreement proportion differed by year and urbanicity, with larger disagreement in the rural category than in micropolitan and metropolitan categories (13%, 7%, 6.2%, respectively).
Discussion and conclusion
We describe a novel pipeline that can facilitate the secure acquisition and sharing of neighborhood-level SDOH without sharing PHI. The pipeline can be scaled to include additional social, climate, and environmental variables, and can be extended to an unlimited number of health systems.
Citation: Schappe T, McElroy LM, Ogundolie M, Matsouaka R, Rogers U, Bhavsar NA (2025) A data pipeline for secure extraction and sharing of social determinants of health. PLoS ONE 20(1): e0317215. https://doi.org/10.1371/journal.pone.0317215
Editor: Celine Rozenblat, University of Lausanne, SWITZERLAND
Received: July 23, 2024; Accepted: December 23, 2024; Published: January 31, 2025
Copyright: © 2025 Schappe et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Data cannot be shared publicly because of ethical restrictions from the Duke University Health System as the data contain potentially identifying and sensitive patient information. Data could however be available for researchers who meet the criteria for access to confidential data, upon reasonable request to the authors and with permission of the Duke University Health System Institutional Review Board, Suite 900 Erwin Square, 2200 West Main Street, Campus Box #104026, Durham, NC 27705; Phone: (919) 668-5111.
Funding: This work was supported by the National Institute on Minority Health and Health Disparities under Award Numbers U54MD012530 and K08MD017632-01 (McElroy); the American Surgical Association Foundation (McElroy); a Harold Amos Medical Faculty Development Award from the Robert Wood Johnson Foundation (McElroy); the National Heart, Lung, and Blood Institute under Award Number HLK01140146 (Bhavsar); and the National Center for Advancing Translational Sciences under Award Number UL1TR002553 (Bhavsar, Matsouaka, Schappe).
Competing interests: The authors have declared that no competing interests exist.
Abbreviations: ADI, Area Deprivation Index; DeGAUSS, Decentralized Geomarker Assessment for Multi-site Studies; DUHS, Duke University Health System; ERS, Economic Research Service; FIPS, Federal Information Processing Standards; PHI, Protected health information; RUCA, Rural-Urban Commuter Area; SDOH, Social determinants of health; TIGER, Topologically Integrated Geographic Encoding and Referencing; USDA, United States Department of Agriculture
Introduction
Social determinants of health (SDOH) are defined by the World Health Organization as “the conditions in which people are born, grow, live, work, and age, along with the wider set of forces and systems shaping the conditions of daily life” [1]. SDOH indicators are commonly assessed at the individual level, through measures like race, income and education level. However, neighborhood-level SDOH can provide additional important information about a patient’s social context, including living conditions, access to transportation, and community support. Assessment of SDOH indicators at multiple levels (e.g. patient, community, institution, etc.) provides an opportunity to examine upstream drivers of individual and population-level health inequities [2].
Incorporating neighborhood-level SDOH into research and quality improvement efforts presents unique challenges relative to other patient data. Within health systems, patient address data requires geocoding to link patients with public data sources that characterize neighborhood conditions, such as environmental quality, safety, or poverty [3]. Multicenter studies face an additional data security challenge because patient address is considered protected health information (PHI); obtaining the permissions and establishing the infrastructure for securely sharing PHI are costly and labor intensive.
Systematic approaches that securely link individual patients to neighborhood-level SDOH data can enable assessment of SDOH within health systems and the communities. It also allows examination of individual and community level SDOH across health systems, enabling the integration of community context into multi-center studies focused on care delivery and community context. Fundamental to these approaches is the need to accurately, reliably, and securely geocode patient addresses. Decentralized Geomarker Assessment for Multi-site Studies (DeGAUSS), is an open-source privacy-preserving software package that geocodes addresses in a standardized manner which can be used across multiple health systems. Yet, the DeGAUSS software suite includes limited area-level SDOH variables, and it lacks an automated pipeline to generate geocoded SDOH variables from address data. The goals of this study were to 1) develop a pipeline for generating community-level SDOH that augments the features provided by DeGAUSS and 2) compare DeGAUSS geocoding results to a vendor tool currently used by the Duke University Health System (DUHS) on a large, geographically diverse patient cohort in order to assess its viability for multi-health system analyses as a secure and standardized alternative to commercial geocoding software. Because geocoding accuracy directly affects downstream SDOH data and since previous work has found systematic differences in accuracy by rurality, we stratified our assessments by both rurality and by Area Deprivation Index (ADI); we also examined accuracy by year to elucidate potential temporal patterns [4].
Materials and methods
Data pipeline creation
Our primary goal was to develop a pipeline for generating community-level SDOH that augments the features provided by DeGAUSS. We created a flexible in-house data pipeline centered around DeGAUSS v3.3.0 that utilizes containerization via Singularity [5] to geocode patient addresses and obtain neighborhood-level SDOH variables. The pipeline requires a comma-separate values (CSV) file of address text as input and provides a CSV file with each address and associated de-identified SDOH variables as output. The pipeline checks for the existence of the required Singularity containers and pulls them from the Duke public container registry if needed. After container verification, the pipeline performs the following steps, checking for successful output after each: 1) normalize input address formatting, 2) geocode address using the 2021 TIGER/Line shapefiles via DeGAUSS, 3) identify census tract and block group level FIPS codes, 4) merge with a curated SDOH data repository, and 5) de-identify output data (Fig 1). The address normalization step standardizes formatting beyond street numbers, such as apartment units or suites. The SDOH data repository is a publicly-available auxiliary resource [6] created by compiling various publicly-available SDOH data sources by spatial scale; at present it contains 77 distinct fields and it can be expanded depending on the research question in future studies. Additional de-identification steps can also be added, including removal of PHI, outlier truncation, and the addition of random noise to continuous variables.
Data sources
Electronic health records.
We obtained address data from the Duke University Health System (DUHS) for a cohort of adult patients referred for abdominal transplant between January 1, 2016 and December 31, 2022. DUHS is a tertiary health system that consists of three hospitals and a network of primary and specialty care clinics. This study was approved by the Duke University IRB; data were accessed on March 1, 2023 and the authors had access to information that could identify individual participants during and after data collection, with IRB permission. The need for consent was waived by the DUHS IRB.
Area deprivation index.
Neighborhood socioeconomic status was quantified with the Area Deprivation Index (ADI). The ADI is a composite score that ranks neighborhoods by socioeconomic disadvantage, based on factors related to employment, income, and education at state and national levels [7]. We assigned quintile categories to each geocoded address based on the 2018 national percentile ranking. Addresses in census units with no ADI value were assigned to a separate ‘missing’ category.
Rural-urban commuting area.
We classified each address into one of five urban/rural categories based on the 2010 Rural-Urban Commuting Area (RUCA) classifications. The 2010 RUCA codes published by the USDA Economic Research Services (ERS) classify census tracts using measures of population density, urbanization, and daily commuting.
Geocoding methods
Duke university health system (DUHS) geocoding.
DUHS uses a third-party vendor geocoding tool to provide census unit assignments from patient addresses. The tool is used for health system-level geocoding of current and historic addresses for DUHS patients.
DeGAUSS geocoding.
DeGAUSS is an open-source software suite that was developed specifically for facilitating geocoding of patient addresses for multi-site studies. It provides secure geocoding functionality that uses a local instance of the TIGER/Line Shapefiles database provided by the US Census Bureau [8]. DeGAUSS assigns census tract and block group level FIPS codes based on address geolocations. The security of DeGAUSS is based on containerization, a system-level virtualization of a pre-built compute environment that includes software tools and databases. Containerization allows workflows to be executed within an isolated compute instance using self-contained tools and databases, preventing exposure of PHI outside of the compute environment. The compute environment is encapsulated in a single file (‘container’) which can easily be run on a variety of computational infrastructures, including secure enclaves designed for handling PHI. In the case of DeGAUSS, the required software along with the U.S. Census TIGER/Line Shapefiles database are built into the container, precluding the need to query remote databases and share PHI to obtain geolocation based SDOH.
Statistical analysis
Our secondary goal was to examine overall agreement between geocoding output from DeGAUSS and the third-party vendor geocoding tool. To accomplish this, we first quantified concordance with respect to census unit assignment (i.e. the census tract and block group Federal Information Processing Standards (FIPS) code assignment based on the address geolocation). We hypothesized that concordance may differ by neighborhood socioeconomic status, by urban-rural category, and by time. A second way in which we characterized agreement was the geographic distance between the geocoded locations provided by each tool. Prior to analysis, we processed the raw geocoding output for the patient cohort address data using DeGAUSS and the vendor tool. We combined the raw DeGAUSS geocoding results with the original address dataset that contained the vendor tool geocoding results. We excluded addresses that were not successfully geocoded by both tools and collapsed duplicative addresses into the first instance the address was reported.
Overall concordance.
We compared the performance of DeGAUSS to the vendor tool by quantifying the concordance of census unit assignments at both the census tract and census block group levels. To do so, we utilized two metrics: Cohen’s kappa [9], a commonly-used index that accounts for agreement that is expected to occur by random chance, and percent disagreement. Because of the large number of categories for classification, we calculated Cohen’s kappa following the generalized equations outlined in Pontius and Millones, 2011 [10]. To examine proportion disagreement, we first derived a binary outcome that indicates whether the census unit assignments differed among the tools. We then fit an intercept-only logistic regression model using this binary outcome to estimate percent disagreement (as the estimated probability) and corresponding 95% confidence intervals for both census tract and block group levels.
Heterogeneity of concordance by strata.
In addition to overall concordance, we quantified differences in concordance of census unit assignments among strata for urban/rural categories, socioeconomic deprivation categories, and the first calendar year that the address was valid for the patient. We calculated the corresponding Cohen’s Kappa by strata and used a Bonferroni correction [11] to account for multiple comparisons in the resulting p-values [12]. We then examined whether percent disagreement differed by strata by fitting a logistic regression model using the same outcome as above except with one stratification variable as a covariate. We used a likelihood ratio test to compare this full model to the previously fit intercept-only model to assess the evidence for differences in percent disagreement among strata. We also used the full model to estimate percent disagreement and corresponding Bonferroni-adjusted 95% confidence intervals for each stratum. In cases with a significant likelihood ratio test (i.e., p < 0.05), we performed follow-up t-tests for all pairwise comparisons among all strata and used Tukey adjustment for the resulting p-values [13].
Geographic distance by strata.
Because distance to transplant center is a commonly studied SDOH, we used geographic distance between geocoded locations provided by each tool for each address as a supplementary metric to compare their performance [14–17]. We first calculated the straight-line geographic distances between geolocations provided by each tool for all addresses [18]. We then calculated median distances and performed non-parametric Kruskal-Wallis tests for differences in the distribution of distances by strata. In the case of a significant Kruskal-Wallis test, we performed Wilcoxon rank sum tests for all pairwise comparisons among strata and used the Bonferroni adjustment method to account for multiple comparisons in the resulting p-values and 95% confidence intervals. Data analysis was performed using the R project for statistical computing v4.3.1 [19].
Results
A total of 13,562 unique addresses associated with patients referred for an abdominal transplant at DUHS from January 1, 2016, to December 31, 2022, were geocoded by the third-party vendor geocoding tool, of which 13,262 (97.8%) DeGAUSS was able to geocode successfully. The population of addresses originated from 42 different U.S. states, with 10,209 (75.3%) being from metropolitan areas, 1,991 (14.7%) from micropolitan locations, 845 (6.2%) from small towns, 239 (1.8%) from rural regions, and 278 (2.0%) missing a RUCA designation. With respect to ADI, 974 (7.2%) addresses were in the least deprived quintile of block groups nationally, 2,328 (17.2%) were in the 2nd quintile, 3,294 (24.3%) were in the 3rd quintile, 3,567 (26.3%) were in the 4th quintile, 3,222 (23.8%) were in the most highly deprived quintile of block groups, and 177 (1.3%) of addresses had an undefined ADI.
Overall, at census block group and census tract levels, there was high agreement as measured by Cohen’s kappa (0.93 and 0.95, respectively) and low percent disagreement (6.7% and 4.6%, respectively) between the vendor tool and DeGAUSS in census unit assignments (Table 1).
The overall percent disagreement between DeGAUSS and the vendor tool differed significantly by urbanicity at the block group level (P < 0.001) but not at the census tract level (P > 0.98, Table 2).
The percent disagreement differed by the calendar year that the address was applicable to the patient at the block group level (P < 0.003) but not at the tract level (P > 0.08, Table 2). There was no heterogeneity in percent disagreement by ADI quintiles at either block group (P > 0.52) or tract (P > 0.99, Table 2) levels.
Concordance by Urban-rural
In analyses stratified by urbanicity, there was greater estimated concordance in geocoding in urban areas as compared to rural areas (Table 3).
Metropolitan areas had a high estimated index of agreement (0.94; 95% CI: [0.93, 0.94]) and a low estimated percent disagreement (6.2%; 95% CI: [5.6%, 6.8%]) at the block group and census tract level (index of agreement: 0.96, 95% CI: [0.95, 0.96]; percent disagreement: 4.3%, 95% CI: [3.9%, 4.9%]). By contrast, rural addresses had a lower estimated index of agreement at the block group and tract levels (0.87 and 0.93, respectively) and a higher estimated percent disagreement (13.0% and 7.1%, respectively), the latter of which was a statistically significant pairwise contrast (P < 0.001, Fig 2A). Moreover, estimated percent disagreement among rural addresses was significantly higher (P < 0.008) than among micropolitan addresses (7.0%, 95% CI: [5.7%, 8.6%]) at the census block group level. Additionally, estimated percent disagreement among ‘small town’ addresses (9.1%, 95% CI: [6.9%, 11.9%]) was significantly higher (P < 0.006) than for metropolitan addresses (6.2%, 95% CI: [5.6%, 6.8%], Fig 2A).
Pairwise comparisons of probability of disagreement of census unit assignments between DeGAUSS and vendor tool geocoding at the block group level among A) urban/rural categories and B) calendar year address was applicable.
Concordance by address year
In analyses stratified by the first calendar year that the address was observed, estimated percent disagreement between the geocoders at the block group level was significantly different between addresses first observed prior to 2016 (6.0% disagreement (95% CI: [5.3%, 6.8%])) as compared to addresses first observed in 2017 (9.0% disagreement (95% CI: [6.6%, 12.3%], P < 0.026, Fig 2B). There was no evidence for heterogeneity in percent disagreement by year at the tract level (P > 0.05, Table 2).
Geographic distance
For most addresses (92%), the distance between geocoded locations provided by each tool was less than 1,000 ft (S1 Table). However, 4% of addresses had distances greater than 1 mile, including 0.95% of metropolitan addresses with distances greater than 10 miles (S1 Fig, S1 Table). The distributions of geographic distances between the geocoding tools were right skewed, as well, within urban-rural categories (S1 Table) and within ADI quintile categories (S2 Table). There were significant differences in the distances between geocoded locations provided by the vendor tool and DeGAUSS by urban/rural categories (P < 0.001, S3 Table). There were significant shifts (i.e. stochastic dominance) among the distributions of geographic distances for all pairwise comparisons among urban/rural categories except for the ‘small town’ vs ‘rural’ comparison (S4 Table). For example, the estimated median difference in geographic distance for rural addresses compared to metropolitan addresses was 93.3 ft. (95% CI: 63.6 ft., 127.7 ft.).
There were significant differences in the distances between geocoded locations provided by the vendor tool and DeGAUSS by ADI quintile (P < 0.014, S2 Table), with the largest median distance in the 4th ADI quintile of 140 ft (S2 Table). Among all pairwise comparisons for stochastic dominance, the only statistically significant result was between the 2nd and 4th quintile groups (P < 0.01), but this is unlikely to be practically meaningful as the estimated median difference in distance between the two quintiles was 8.79 ft. (95% CI: [3.62 ft., 14.52 ft.]) (S5 Table).
Discussion
Neighborhood-level SDOH are increasingly recognized as upstream drivers of health disparities [20–22], but can be challenging to include in analyses of EHR data because of privacy concerns. For projects that include multiple health systems, generating and sharing neighborhood SDOH requires either expertise with geocoding at each site and ensuring geocoding practices are the same across sites, or designating one site as a coordinating center, obtaining IRB approval and creating data use agreements that allow sharing of PHI. There is a need to develop reproducible privacy-preserving methods that allow for incorporation of community level SDOH into multicenter research without the need to share PHI [23–26].
We created a generalized data pipeline using DeGAUSS to provide consistent geocoding and generation of neighborhood based SDOH data for sharing across health systems without sharing PHI. These features can facilitate multi-health system participation of large-scale policy relevant research, such as the impact of climate on health or state subsidies for areas lacking referring providers, by allowing the inclusion of geographically diverse populations. The pipeline can be shared with an unlimited number of health systems and its flexibility enables future extensions to be introduced. Additional SDOH variables can be included and easily pushed to participating centers, including raster data such as Normalized Difference Vegetation Index (NDVI); the complexity of this process depends on the variable source and type. Further, customized definitions of neighborhoods can be utilized assuming corresponding polygons exist. We note that ensuring correct temporality of exposures relative to health outcomes is not ensured by DeGAUSS or our pipeline, but rather falls to the investigator or analyst.
We assessed the accuracy of DeGAUSS geocoding by comparing the resulting census tract and block group assignments to those provided by the third-party vendor geocoding tool available at DUHS. DeGAUSS was able to geocode and provide census assignments for 97.8% and demonstrated higher agreement with a third-party vendor geocoding tool at the census tract level compared to the block group level. There was no heterogeneity in percent disagreement among ADI quintiles at either tract or block group levels, indicating that DeGAUSS can be used effectively across socioeconomic strata. However, we found a higher level of disagreement for rural addresses compared with both micropolitan and metropolitan categories, and “small-town” addresses as compared to metropolitan addresses. This was supported by geographic distances between geocoded locations, which were larger for rural and small-town addresses compared to suburban and urban areas. This is not unexpected given differences in geographic scales and mail delivery infrastructure such as rural route numbers, and consistent with prior research demonstrating a relationship between rurality and geocoding ability [27]. Higher disagreement in geocoding tools may pose a challenge when examining urban/rural disparities. There is a tradeoff between greater address inclusion, which can decrease selection bias, and decreased geocoding accuracy, which increases information bias. Future work should examine more advanced methods of characterizing social context within rural geographic regions.
We observed moderate discordance between DeGAUSS and the third-party vendor geocoding tool in urban and suburban areas; surprisingly, nearly 100 metropolitan addresses had distances of greater than 10 miles between the geocoded locations. This may be indicative of geocoding failure for at least one of the tools, but without a validated gold standard to compare against, it is not possible to determine which tool is correct. One potential explanation could be an anti-conservative string search algorithm used to query the input addresses given the absence of a perfect match. For example, it’s possible that one tool is matching addresses not included in its database to an incorrect address in a different city or town that shares common features with the query, such as a street name. While this geocoding approach has the potential to result in large error rates, only a small fraction of addresses in our cohort were affected.
There are some limitations to note. First, there is no definitive way to determine which geocoding tool (or neither) provides the true correct geocoded location. We hypothesize that short distances between geocoded addresses is small (e.g., 50 feet), indicate high accuracy in geocoding [12,28]. Second, there are several products on the market to facilitate geocoding of patient address data. We compared DeGAUSS to the DUHS third-party geocoding tool with good results, but performance may vary across other products and software.
Conclusion
Incorporating neighborhood-level SDOH into research and quality improvement efforts can identify upstream drivers of disparities in care ripe for intervention. Unfortunately, SDOH data are housed across multiple national data sources and organized at varying geographic levels. We describe a novel pipeline that is driven by DeGAUSS, an open-source software application capable of geocoding patient addresses and SDOH data without the need to share PHI. We found that DeGAUSS can address these concerns and is an accurate, open source, privacy-preserving tool that can facilitate incorporation of community level SDOH into multicenter studies that use EHR data.
Supporting information
S1 Fig. Number of addresses by distance (mi.) between geocoded locations stratified by urban-rural category for addresses with distances of greater than 2 miles.
https://doi.org/10.1371/journal.pone.0317215.s001
(DOCX)
S1 Table. Percentage of addresses by geographic distance between geocoded locations provided by DeGAUSS and the vendor tool geocoder, stratified by urban-rural category.
https://doi.org/10.1371/journal.pone.0317215.s002
(DOCX)
S2 Table. Distance (ft.) between geocoded locations provided by DeGAUSS and vendor tool, stratified by quintiles of Area Deprivation Index.
https://doi.org/10.1371/journal.pone.0317215.s003
(DOCX)
S3 Table. Distance (ft.) between geocoded locations provided by DeGAUSS and vendor tool, stratified by urban-rural category.
https://doi.org/10.1371/journal.pone.0317215.s004
(DOCX)
S4 Table. Estimated median difference in distance between samples of addresses drawn from each group in the comparison by urban-rural category.
https://doi.org/10.1371/journal.pone.0317215.s005
(DOCX)
S5 Table. Estimated median differences in distance between samples of addresses drawn from each group in the comparison by quintile of Area Deprivation Index (ADI).
https://doi.org/10.1371/journal.pone.0317215.s006
(DOCX)
References
- 1.
Health CoSDo. Closing the gap in a generation: health equity through action on the social determinants of health: final report of the commission on social determinants of health: World Health Organization; 2008.
- 2. Paskett E, Thompson B, Ammerman AS, Ortega AN, Marsteller J, Richardson D. Multilevel Interventions To Address Health Disparities Show Promise In Improving Population Health. Health Aff (Millwood). 2016;35(8):1429–34. pmid:27503968; PubMed Central PMCID: PMC5553289.
- 3. Lindenfeld Z, Pagan JA, Chang JE. Utilizing Publicly Available Community Data to Address Social Determinants of Health: A Compendium of Data Sources. Inquiry. 2023;60:469580231152318. pmid:36803137; PubMed Central PMCID: PMC9940168.
- 4. Chow TE, Dede-Bamfo N., & Dahal K. R. Geographic disparity of positional errors and matching rate of residential addresses among geocoding solutions. Annals of GIS. 2015;22(1):29–42.
- 5. Kurtzer GM, Sochat V, Bauer MW. Singularity: Scientific containers for mobility of compute. PLoS One. 2017;12(5):e0177459. Epub 20170511. pmid:28494014; PubMed Central PMCID: PMC5426675.
- 6. Schappe TM, Lisa M.; Matsouaka Roland; Bhavsar Nrupen A. SDOH Data Repository 2023 [cited 2023]. Available from: https://chart-consortium.pages.oit.duke.edu/sdoh-data-repository/.
- 7. Kind AJH, Buckingham WR. Making Neighborhood-Disadvantage Metrics Accessible—The Neighborhood Atlas. N Engl J Med. 2018;378(26):2456–8. Epub 2018/06/28. pmid:29949490; PubMed Central PMCID: PMC6051533.
- 8. Brokamp C. DeGAUSS: Decentralized Geomarker Assessment for Multi-Site Studies. Journal of Open Source Software. 2018;3:812.
- 9. McHugh ML. Interrater reliability: the kappa statistic. Biochem Med (Zagreb). 2012;22(3):276–82. pmid:23092060; PubMed Central PMCID: PMC3900052.
- 10. Pontius R, Millones M. Death to Kappa: Birth of quantity disagreement and allocation disagreement for accuracy assessment. International Journal of Remote Sensing. 2011;32:4407–29.
- 11. CE B. Teoria statistica delle classi e calcolo delle probabilità. Pubblicazioni del R Istituto Superiore di Scienze Economiche e Commerciali di Firenze. 1936;8:3–62.
- 12. Sen A, Li P, Ye W, Franzblau A. Bayesian inference of dependent kappa for binary ratings. Stat Med. 2021;40(26):5947–60. Epub 20210920. pmid:34542193.
- 13.
R L. emmeans: Estimated Marginal Means, aka Least-Squares Means. R package version 1.8.6. https://github.com/rvlenth/emmeans2023.
- 14. Garner L. Distance from a Transplant Center and Getting Listed for a Transplant. Clin J Am Soc Nephrol. 2020;15(4):439–40. Epub 20200324. pmid:32273260; PubMed Central PMCID: PMC7133137.
- 15. McPherson LJ, Barry V, Yackley J, Gander JC, Pastan SO, Plantinga LC, et al. Distance to Kidney Transplant Center and Access to Early Steps in the Kidney Transplantation Process in the Southeastern United States. Clin J Am Soc Nephrol. 2020;15(4):539–49. Epub 20200324. pmid:32209583; PubMed Central PMCID: PMC7133136.
- 16. Schucht J, Davis EG, Jones CM, Cannon RM. Does Distance to Transplant Center Affect Post Kidney Transplant Readmission Rates? Am Surg. 2019;85(8):834–9. pmid:32051063.
- 17. Webb GJ, Hodson J, Chauhan A, O’Grady J, Neuberger JM, Hirschfield GM, et al. Proximity to transplant center and outcome among liver transplant patients. Am J Transplant. 2019;19(1):208–20. Epub 20180803. pmid:29981195; PubMed Central PMCID: PMC6491997.
- 18. Pebesma E. Simple Features for R: Standardized Support for Spatial Vector Data. R Journal. 2018;10:439–46.
- 19.
Team RC. R: A Langauge and Environment for Statistical Computing. 4.3.1 ed. Vienna, Austria: R Foundation for Statistical Computing; 2023.
- 20. Phelan JC, Link BG, Tehranifar P. Social conditions as fundamental causes of health inequalities: theory, evidence, and policy implications. J Health Soc Behav. 2010;51 Suppl:S28–40. pmid:20943581.
- 21. Link BG, Phelan J. Social conditions as fundamental causes of disease. J Health Soc Behav. 1995;Spec No:80–94. pmid:7560851.
- 22.
In: Baciu A, Negussie Y, Geller A, Weinstein JN, editors. Communities in Action: Pathways to Health Equity. Washington (DC)2017.
- 23. Lemke D, Mattauch V, Heidinger O, Hense HW. [Who Hits the Mark? A Comparative Study of the Free Geocoding Services of Google and OpenStreetMap]. Gesundheitswesen. 2015;77(8–9):e160–5. Epub 20150708. pmid:26154258.
- 24. Zandbergen P. A comparison of address point, parcel and street geocoding techniques. Computers, Environment and Urban Systems. 2008;32:214–32.
- 25. Jacquemin B, Lepeule J, Boudier A, Arnould C, Benmerad M, Chappaz C, et al. Impact of geocoding methods on associations between long-term exposure to urban air pollution and lung function. Environ Health Perspect. 2013;121(9):1054–60. Epub 20130703. pmid:23823697; PubMed Central PMCID: PMC3764075.
- 26. Marsolo K, Kiernan D, Toh S, Phua J, Louzao D, Haynes K, et al. Assessing the impact of privacy-preserving record linkage on record overlap and patient demographic and clinical characteristics in PCORnet(R), the National Patient-Centered Clinical Research Network. J Am Med Inform Assoc. 2023;30(3):447–55. pmid:36451264; PubMed Central PMCID: PMC9933062.
- 27. Brokamp C, Wolfe C, Lingren T, Harley J, Ryan P. Decentralized and reproducible geocoding and characterization of community and environmental exposures for multisite studies. J Am Med Inform Assoc. 2018;25(3):309–14. pmid:29126118; PubMed Central PMCID: PMC7378876.
- 28. Basu S, Banerjee M, Sen A. Bayesian inference for kappa from single and multiple studies. Biometrics. 2000;56(2):577–82. pmid:10877319.